Adder device, data accumulation method and data processing device

Abstract

The invention discloses an adder device, a data accumulation method, and a data processing device. The adder device includes: a first adder module, an addition tree unit composed of a multi-stage adder array and a first control unit, the addition tree The unit accumulates data step by step based on the control signal of the first control unit; the second adder module, including a two-input addition and subtraction operation unit and a second control unit, performs addition or subtraction on the input data; the shift operation module , used to perform a left shift operation on the output data of the first adder module; the AND operation module is used to perform an AND operation on the output data of the shift operation module and the output data of the second adder module; the controller module uses It controls the data input of the first adder module and the second adder module, controls the shift operation of the shift operation module, and controls the emission of the control signal of the first control unit and the second control unit. Thus, fast accumulation of data is realized.

Description

technical field [0001] The invention belongs to the field of data processing systems, and relates to data processing of an adder device, in particular to an adder device, a data accumulation method and a data processing device including the adder device. Background technique [0002] In recent years, neural network algorithms based on the M-P neuron model have been widely used in product recommendation, image recognition and other fields. The M-P neuron model is an additive neuron model. The output value of the neuron is equal to the weighted sum of the neuron input and the result of nonlinear transformation after accumulating the bias value. That is, a large number of accumulation operations and addition operations are required in the neural network algorithm. At the same time, in the training process of neural network computing, not only a large number of accumulation operations are included, but also subtraction operations are included. [0003] When using traditional g...

AI Technical Summary

Problems solved by technology

[0003] When using traditional general-purpose processing processors to operate neural network algorithms or train a neural network, only two data can be added each time, which is inefficient.

And if fixed-point data is used for operations, it is necessary to deal with the overflow of addition

Images